"Fake Frames" Tested | DLSS 4.0, MFG 4X, & NVIDIA's Misleading Review Guide

00:38:58
https://www.youtube.com/watch?v=Nh1FHR9fkJk

Ringkasan

TLDRSelles videos arutatakse Nvidia multiframe generation (MFG) tehnoloogiat DLSS 4.0 raames, tuues esile mitmeid probleemseid kohti, sealhulgas vananenud lähenemisviise ja koostöö puudumist avatud allika tarkvaraga. Video käsitleb MFG eeliseid ja väljakutseid, rõhutades turunduse ja esituse segadust. Samuti uuritakse pildikvaliteedi täiustusi ja jõudluse parandusi, tuues esile võimalikke vale arusaamu frame generation'i eesmärgist. Video kutsub üles selgele suhtlemisele ja koostööle jõudluse mõõtmise tööriistade osas.

Takeaways

  • 🎮 Nvidia on multiframe generation (MFG) tehnoloogia kesksel kohal.
  • 🤖 DLSS 4.0 introduces new features for improved gaming.
  • 📉 Nvidia kasutab vananenud avatud allika tööriistu, mis tekitab segadust.
  • 🔄 MFG ei paranda madala fps jõudlusi, vaid sujuvust.
  • 🔍 Pildikvaliteedi täiustused toovad kaasa muret jõudluse vähenemise kohta.
  • 📊 Uus pilditöötluse mudel (Transformer) lubab paremat stabiilsust.
  • 💼 Oluline on avatud allika tarkvara kõigi osaliste koostöö.
  • 🚀 Nvidia peab paremini suhtlema avatud allika kogukonnaga.

Garis waktu

  • 00:00:00 - 00:05:00

    Nvidia'i suhtumine avatud lähtekoodiga tarkvarasse ja multiframe genereerimise (MFG) tehnoloogiasse tekitab muredekarusselli. Nvidia on lisanud avatud lähtekoodiga tarkvara oma Frame View tööriistasse, tehes muudatusi, mis näivad olevat vananenud ja eksitavad. Loodud arusaama kohaselt ei ole Nvidia teadlikud oma tarkvara aegumusest ning nad on läinud iseseid teed, mis kahjustab mõõtmistulemusi.

  • 00:05:00 - 00:10:00

    Rääkides multiframe genereerimisest, jagame uuringud DLSS 4.0 süsteemi tutvustamiseks, millel on täiustused võrreldes varasemate versioonidega. DLSS 4.0 fookus on frame generation'il ja kuigi on teatud arusaamad, millega nõustume, peavad kõik osalised arvestama oluliste mõõtmistermide ja -meetoditega.

  • 00:10:00 - 00:15:00

    DLSS, nagu selgitab Patrick, on süvamugav õppe superproovide tehnoloogia, mida on arendatud ja ajakohastatud. Nvidia on liikunud uute mudelite suunas, et saavutada paremat kvaliteeti, kuid alles jäävad varem komplitseeritud ja sageli segadust tekitavad määratlused ja numeratsioonid.

  • 00:15:00 - 00:20:00

    Uus Optical Multiframe Generation (OMFG) omadus, samuti DLSS Super Resolution'i renaming on osa Nvidia's sagedaste ümbernimetamisprotsessidest. Need muudatused võimaldavad Nvidia-l täiustada oma Geneerimise võimet, pakkudes samas segaseid teadmisi tarbijatele, millega nad saavad kokku puutuda.

  • 00:20:00 - 00:25:00

    Tulevaste funktsioonide, sealhulgas raamistiku hindamine, vajadus rõhutada latentsuse vähendamist ning uute tehnoloogiate toimetulekut on põhiteema. Kuigi Nvidia väidab, et nad suurendavad jõudlust ja stabiilsust, peavad testid ilmselt sellele vastu panema.

  • 00:25:00 - 00:30:00

    Nvidia'i Frame View tööriista palee, sealjuures avaldatud andmed ja vabatahtlikud koostööd Intel'iga näitavad, et üleminek avatud lähtekoodiga tarkvara pole olnud ühesuunaline koostöö, vaid pigem individuaalsed ja isolatsioonis tehtud muudatused.

  • 00:30:00 - 00:38:58

    Viimase osana tuuakse välja mõõtmis- ja katsetamismetoodika keerukus. Nvidia peab olema ettevaatlik oma turundusslaidide ja müügi väidete sidumisel ning tagama, et kasutajad mõistavad lahenduse tõhusust ja piire. MFG ei tohiks olla lahendus madala framerate'i olukordadele, vaid pigem pehme ja sujuva mängukogemuse rõhutamine juba aktiivsete frame rate'ide väärtustega.

Tampilkan lebih banyak

Peta Pikiran

Video Tanya Jawab

  • What is DLSS 4.0?

    DLSS 4.0 stands for Deep Learning Super Sampling, an updated feature set that includes Multiframe Generation (MFG).

  • What problems does Nvidia face regarding open source software?

    Nvidia has faced criticism for using outdated versions of open-source tools and not collaborating effectively with developers like Intel, leading to confusion and measurement inconsistencies.

  • How does MFG work in DLSS 4.0?

    Multiframe Generation involves generating additional AI frames to improve performance by interpolating frames between the current and previous rendered frames.

  • What are the performance implications of using MFG?

    MFG aims to enhance already playable frame rates rather than improve low frame rates, providing smoother experience without necessarily increasing the base frame rate.

  • What is the new image enhancement model introduced in DLSS 4.0?

    Nvidia is switching to a Transformer model for image processing, claiming improved stability and reduced ghosting, though it may slightly reduce performance.

Lihat lebih banyak ringkasan video

Dapatkan akses instan ke ringkasan video YouTube gratis yang didukung oleh AI!
Teks
en
Gulir Otomatis:
  • 00:00:00
    [Music]
  • 00:00:01
    as simply as possible Nvidia is being a
  • 00:00:04
    dick Nvidia wants everyone to talk about
  • 00:00:05
    multiframe generation we're happy to do
  • 00:00:07
    that because it is worth talking about
  • 00:00:09
    but we can't talk about MFG which is
  • 00:00:11
    part of dss4 now without also talking
  • 00:00:14
    about all of the problems that envidia
  • 00:00:16
    has been leaving in its wake and its
  • 00:00:18
    misunderstanding of how open- Source uh
  • 00:00:21
    present mod software even works right
  • 00:00:22
    now so Nvidia has inducted open source
  • 00:00:24
    software into its frame view tool it's
  • 00:00:27
    made changes and what it thinks are
  • 00:00:28
    improvements to those tools for its own
  • 00:00:31
    frame view software without as far as we
  • 00:00:33
    can tell
  • 00:00:47
    recontruction with Nvidia back at CES
  • 00:00:49
    when they were announcing these cards
  • 00:00:51
    they didn't seem aware at all of what
  • 00:00:53
    was outdated about their understanding
  • 00:00:55
    and the v1.1 version of the review guide
  • 00:00:58
    does technically add in some notes about
  • 00:01:00
    CPU busy and CPU weight which we're
  • 00:01:02
    going to talk about today but it doesn't
  • 00:01:04
    change the fact that some of the
  • 00:01:06
    foundational information in there uh
  • 00:01:09
    appears to be just out ofd and it's
  • 00:01:11
    because they decided to do their own
  • 00:01:13
    thing and use what looks like a year old
  • 00:01:15
    version of preson as the framework for
  • 00:01:18
    frame view rather than just working with
  • 00:01:20
    everyone else while Nvidia has some
  • 00:01:22
    valid points in how to measure generated
  • 00:01:24
    frames the biggest Counterpoint is that
  • 00:01:26
    their statements were just largely based
  • 00:01:28
    on old outdated software that's already
  • 00:01:31
    made a shift away from the MS between
  • 00:01:33
    presents metric that Nvidia is now
  • 00:01:34
    contesting and yes they mentioned CPU
  • 00:01:36
    busy and wait but we just feel like the
  • 00:01:38
    entire premise of how this was presented
  • 00:01:40
    was misleading the open source license
  • 00:01:41
    that present Mon's filed under allows
  • 00:01:43
    Nvidia to do this it's just not a
  • 00:01:44
    productive thing to do and we think it's
  • 00:01:46
    potentially detrimental to testing
  • 00:01:47
    effort so uh Intel has told us that
  • 00:01:50
    Nvidia is welcome to contribute to
  • 00:01:51
    present Monon Which Intel maintains and
  • 00:01:53
    play nice with everyone else but Nvidia
  • 00:01:55
    is doing Nvidia things as for why this
  • 00:01:57
    is happening again this is all relevant
  • 00:01:58
    because multiframe generation which is
  • 00:01:59
    the topic we're diving into today
  • 00:02:01
    whatever you want to call it generated
  • 00:02:02
    frame uh artificial frame generated
  • 00:02:05
    frame or they'll love this one fake
  • 00:02:07
    frame whatever word you use they are
  • 00:02:09
    very difficult to test and measure
  • 00:02:10
    accurately and we're only scratching the
  • 00:02:13
    surface of this topic today there's a
  • 00:02:14
    whole lot of other stuff to do like
  • 00:02:15
    image quality in the future this video
  • 00:02:17
    will first Define what dlss 4.0 is and
  • 00:02:19
    some of the feature set differences
  • 00:02:20
    between that and previous dlss versions
  • 00:02:23
    we'll also go into uh nvidia's framing
  • 00:02:26
    of the benchmarking methodology the
  • 00:02:28
    things we agree with some of the stuff
  • 00:02:29
    they say is absolutely right and the
  • 00:02:30
    things we disagree with and additionally
  • 00:02:32
    we'll talk about the frame generation
  • 00:02:33
    models and how they perform we do think
  • 00:02:35
    there's something interesting here it's
  • 00:02:36
    absolutely worth talking about uh but
  • 00:02:38
    it's going to take a couple videos to
  • 00:02:39
    really get through it all cuz it's
  • 00:02:41
    complicated before that this video is
  • 00:02:42
    brought to you by the antech flux series
  • 00:02:44
    of cases we previously benchmarked and
  • 00:02:47
    independently reviewed the flux Pro case
  • 00:02:48
    and found it to be one of the best cases
  • 00:02:50
    we tested last year the flux Pro
  • 00:02:52
    features the Modern Trend of wood
  • 00:02:54
    accents while also focusing on a high
  • 00:02:56
    airflow design we found its performance
  • 00:02:57
    to be among the best in our thermal
  • 00:02:59
    testing the case also has a number of
  • 00:03:01
    ease of installation features to make
  • 00:03:03
    assembly and kale management easier
  • 00:03:05
    learn more at the link in the
  • 00:03:06
    description below so despite the
  • 00:03:07
    problems we have with some of nvidia's
  • 00:03:08
    present Monon Framing and some of the
  • 00:03:10
    agreements we have with them dlss 4.0 is
  • 00:03:12
    worth diving into Nvidia seems aware of
  • 00:03:14
    how confusing the roll out of dlss has
  • 00:03:17
    been and continues to be uh and it's mix
  • 00:03:19
    of dlss version numbering plus
  • 00:03:22
    differences in sort of this Matrix of GV
  • 00:03:25
    support versus dlss version support
  • 00:03:27
    that's what's contributed to this so
  • 00:03:28
    we're going to try and break this down
  • 00:03:29
    you may have seen this chart before that
  • 00:03:31
    details the feature set and what does
  • 00:03:33
    and doesn't work across various RTX
  • 00:03:36
    series gpus some of this is segmentation
  • 00:03:38
    some of it is Hardware some of it is
  • 00:03:40
    potentially temporary segmentation that
  • 00:03:42
    could change the future and because this
  • 00:03:44
    topic is so complicated Patrick and I
  • 00:03:45
    are going to split this one up so he's
  • 00:03:47
    going to help with some of the explainer
  • 00:03:48
    and hosting responsibilities on this he
  • 00:03:50
    did the foundational work on writing
  • 00:03:51
    this piece uh and has been researching
  • 00:03:53
    dlss in the different versions for uh
  • 00:03:56
    over a week now so I want to pass it
  • 00:03:57
    over to him to get us started with this
  • 00:04:00
    topic taking a step back dlss stands for
  • 00:04:03
    deep learning super sampling that name
  • 00:04:05
    was already strange seven years ago when
  • 00:04:07
    it was introduced because super sampling
  • 00:04:09
    generally means rendering at a
  • 00:04:11
    resolution above the native screen size
  • 00:04:13
    and then downsampling it for example if
  • 00:04:16
    you rendered at 4K but you have a 1080
  • 00:04:19
    screen you would get a really crisp
  • 00:04:21
    image with excellent
  • 00:04:23
    anti-aliasing uh dlss is kind of the
  • 00:04:26
    opposite for dlss NVIDIA trained an
  • 00:04:29
    algor gthm on a set of extremely high
  • 00:04:31
    resolution uh screenshots on a per game
  • 00:04:34
    basis and what this meant is that they
  • 00:04:36
    could do upscaling in real time for
  • 00:04:39
    example you could render a game at 1080
  • 00:04:41
    on a 4K screen and it would be upscaled
  • 00:04:43
    into a mostly okay looking image dlss is
  • 00:04:46
    no longer trained per game but the name
  • 00:04:48
    is stuck around the use of those
  • 00:04:50
    training images those really
  • 00:04:52
    highresolution screenshots is what makes
  • 00:04:54
    it kind of sort of super sampling all of
  • 00:04:57
    this was pretty straightforward until
  • 00:04:59
    the release of of dlss 3.0 in
  • 00:05:02
    2022 which introduced Optical multiframe
  • 00:05:05
    generation OMFG continuing that video's
  • 00:05:08
    propensity to use initialisms that
  • 00:05:11
    contain other initialisms like uh mfaa
  • 00:05:14
    that was a fun one OMFG is Optical
  • 00:05:16
    multiframe generation later referred to
  • 00:05:18
    as frame generation 2X or just plain
  • 00:05:21
    frame generation because now dlss 4 has
  • 00:05:25
    multiframe generation now with this the
  • 00:05:27
    original functionality of dlss was
  • 00:05:29
    renamed to dlss super resolution and
  • 00:05:32
    that means that Nvidia is now at a point
  • 00:05:34
    where they can recommend deep learned
  • 00:05:37
    super sampling super resolution Ultra
  • 00:05:41
    performance super ultra I added the last
  • 00:05:44
    part as for frame generation frame
  • 00:05:47
    generation falls under the dlss umbrella
  • 00:05:49
    even though it's not directly tied to
  • 00:05:51
    upscaling or super sampling uh dlss 3.5
  • 00:05:56
    added Ray reconstruction to the dlss
  • 00:05:58
    family and now dlss 4 adds multiframe
  • 00:06:01
    generation 3x and 4x as well as changing
  • 00:06:04
    the underlying model for the other dlss
  • 00:06:06
    features so dlss 4 dlss 3.5 dlss 3 at
  • 00:06:11
    Alia are groups of features and updates
  • 00:06:14
    if a game has dlss 4 that means it has
  • 00:06:17
    some of the dlss 4 features from this
  • 00:06:19
    chart most likely multiframe generation
  • 00:06:22
    but not necessarily all of them the
  • 00:06:24
    claims that with 50 Series AI handles 15
  • 00:06:26
    out of every 16 pixels to multiply frame
  • 00:06:29
    rates by up to 8X are based on using the
  • 00:06:32
    maximum 4X MFG setting with the dlss
  • 00:06:35
    super resolution set to Performance
  • 00:06:37
    which was formerly the minimum quality
  • 00:06:39
    setting that would mean that at 4K one
  • 00:06:42
    1080 frame would be rendered for every
  • 00:06:44
    four frames displayed as for MFG MFG is
  • 00:06:48
    available for 50 Series cards and as of
  • 00:06:50
    writing this no other cards support
  • 00:06:52
    multi frame Generation Now frame
  • 00:06:54
    generation can be supported elsewhere
  • 00:06:56
    but MFG is unique right now to the 50
  • 00:06:58
    Series but it's also the centerpiece for
  • 00:07:01
    nvidia's misleading marketing slides
  • 00:07:03
    where it depicted the 90 as doing two
  • 00:07:05
    times the performance of the 4090 this
  • 00:07:08
    misleading chart that we've already
  • 00:07:09
    talked about in the review is because
  • 00:07:11
    MFG includes 4X which is up to three AI
  • 00:07:15
    frames for every one rendered frame MFG
  • 00:07:17
    also includes 3x or up to two AI frames
  • 00:07:20
    for one rendered frame the existing
  • 00:07:22
    frame generation 2x mode has been
  • 00:07:24
    improved Nvidia says but it remains
  • 00:07:26
    exclusive to the 40 series and later and
  • 00:07:29
    frame generation in general still only
  • 00:07:31
    works with games that explicitly
  • 00:07:32
    implement it to paraphrase Nvidia the
  • 00:07:34
    original implementation of frame
  • 00:07:36
    generation combined four elements a
  • 00:07:38
    previous frame a a current frame B an
  • 00:07:42
    quote Optical flow field generated by it
  • 00:07:44
    is Optical flow accelerator end quote
  • 00:07:47
    and data provided directly from the game
  • 00:07:49
    engine about motion and other elements
  • 00:07:51
    using these inputs the card could throw
  • 00:07:53
    together a so-called fake frame or a
  • 00:07:56
    generated one between A and B at the
  • 00:07:58
    cost of up frame B for a little while
  • 00:08:01
    now of course it'd be ridiculous and a
  • 00:08:03
    fundamental misrepresentation of frame
  • 00:08:05
    generation if you were to get up on
  • 00:08:07
    stage and say something untrue like I
  • 00:08:10
    don't know it can predict the future it
  • 00:08:12
    can predict the future uh okay well
  • 00:08:15
    that's beside the point I wonder if this
  • 00:08:17
    can play crisis why why does he do this
  • 00:08:20
    uh it can't extrapolate into the future
  • 00:08:23
    it can interpolate intelligently into
  • 00:08:26
    the past that's the distinction here and
  • 00:08:28
    this gets confusing if you think about
  • 00:08:29
    it too much this creates the feeling of
  • 00:08:31
    a smoother frame rate with some added
  • 00:08:33
    latency which is why Nvidia reflex has
  • 00:08:35
    always been a mandatory setting when
  • 00:08:37
    enabling frame generation a feature that
  • 00:08:39
    we've covered in detail reflex clears
  • 00:08:41
    the render queue so that frames are
  • 00:08:43
    delivered with as little latency as
  • 00:08:45
    possible which helps mitigate the
  • 00:08:47
    latency inherent to frame generation we
  • 00:08:49
    have a full interview into the
  • 00:08:50
    technicals of latency just in general
  • 00:08:52
    including reflex if you'd like to learn
  • 00:08:54
    more so fundamentally we get into one of
  • 00:08:56
    the key areas that this can be kind of
  • 00:08:58
    confusing to follow which is that uh
  • 00:09:00
    frame generation in general is not
  • 00:09:02
    intended it seems to turn low frame
  • 00:09:04
    rates into playable frame rates it seems
  • 00:09:07
    instead that it's more like it's meant
  • 00:09:08
    for turning playable frame rates into
  • 00:09:11
    high frame rates and this is still true
  • 00:09:13
    for MFG as Nvidia stated to digital
  • 00:09:15
    Foundry uh acceptable input frame rate
  • 00:09:19
    is still about the same for 3x or 4X as
  • 00:09:21
    it was for 2x higher input rates means
  • 00:09:23
    fresher data for frame generation
  • 00:09:25
    smaller gaps to bridge and then likewise
  • 00:09:28
    less time to notice flaws in it so
  • 00:09:30
    that's kind of the theory of how it's
  • 00:09:31
    set up uh MFG is still Bridging the same
  • 00:09:35
    gap between the newest rendered frame
  • 00:09:37
    though and the previous one it's just
  • 00:09:40
    doing so with up to three AI frames or
  • 00:09:44
    generated frames or if you want to use a
  • 00:09:46
    term that Nvidia probably hates fake
  • 00:09:48
    frames and it's doing it with up to that
  • 00:09:49
    many instead of one on the plus side
  • 00:09:52
    this theoretically means that latency
  • 00:09:53
    shouldn't be worse with MFG than with
  • 00:09:56
    regular frame generation now to clear up
  • 00:09:58
    another maybe misconception with this if
  • 00:10:00
    you're on a 60 HZ monitor it would
  • 00:10:02
    appear that you are not the target
  • 00:10:04
    audience here for this solution there's
  • 00:10:06
    no point in using frame generation to
  • 00:10:09
    boost your frame rate past the refresh
  • 00:10:11
    rate of the display and that's not just
  • 00:10:13
    our opinion rendering at rates that
  • 00:10:16
    exceed the display refresh rate can
  • 00:10:18
    reduce latency but generating frames at
  • 00:10:21
    a higher rate is just wasting compute
  • 00:10:24
    power on frames you won't see this is
  • 00:10:25
    why nvidia's dss4 demos have topped out
  • 00:10:28
    around 200 40 FPS as for the optical
  • 00:10:31
    flow accelerator we'll throw to one of
  • 00:10:33
    nvidia's public videos that it's
  • 00:10:35
    uploaded where uh it has its own
  • 00:10:37
    resident Steve introduced the technology
  • 00:10:40
    his name is Brian so you may be asking
  • 00:10:42
    what is the optical flow accelerator and
  • 00:10:44
    why do we need it and yes I have met him
  • 00:10:46
    at events and yes we didn't talk about
  • 00:10:49
    on video we we talked about how people
  • 00:10:52
    comment about our hair anyway with dss4
  • 00:10:55
    Ada's Hardware Optical flow accelerator
  • 00:10:58
    is out of the picture dss4 frame
  • 00:11:00
    generation is solely handled by regular
  • 00:11:02
    tensor cores which is why digital
  • 00:11:04
    Foundry asked that question about frame
  • 00:11:06
    generation not MFG potentially come into
  • 00:11:09
    more cards Nvidia claims that the new
  • 00:11:11
    model is faster and uses less vram which
  • 00:11:13
    makes it feel like the only point of the
  • 00:11:15
    OFA was to make frame gen exclusive to
  • 00:11:18
    the 40 series MFG is not possible using
  • 00:11:21
    the old model since it quote would be
  • 00:11:23
    required for every new generated frame
  • 00:11:25
    and the performance cost would throttle
  • 00:11:26
    the GPU resulting in lower input frame R
  • 00:11:29
    end quote Nvidia did confirm with us
  • 00:11:31
    that Blackwell which is the new
  • 00:11:33
    architecture for the 50 Series does have
  • 00:11:35
    a hardware OFA unit we assume here that
  • 00:11:39
    what they're planning to use that for is
  • 00:11:41
    situations where games don't implement
  • 00:11:43
    the new models and maintaining
  • 00:11:45
    compatibility with games that don't have
  • 00:11:47
    that updated frame gen model so uh time
  • 00:11:49
    to get into flip metering and some of
  • 00:11:51
    the issues we mentioned earlier about
  • 00:11:54
    Nvidia doing Nvidia things and uh and
  • 00:11:58
    it's flawed under understanding of
  • 00:11:59
    present Manon so we're going to get into
  • 00:12:01
    that this is also nvidia's stated reason
  • 00:12:03
    for MFG being exclusive to the 50 Series
  • 00:12:06
    Nvidia has given two reasons for locking
  • 00:12:08
    MFG to 50 Series first the fifth
  • 00:12:10
    generation tensor cores are
  • 00:12:12
    significantly improved and second they
  • 00:12:13
    say Blackwell has Hardware support for
  • 00:12:15
    flip metering flip metering shifts
  • 00:12:18
    responsibility for frame pacing from the
  • 00:12:20
    CPU to the GPU driver in order to
  • 00:12:22
    effectively Pace out fake frames or
  • 00:12:24
    generated ones which has become more
  • 00:12:26
    important with MFG this takes place
  • 00:12:29
    immediately prior to scan out again
  • 00:12:31
    Blackwell's implementation of flip
  • 00:12:33
    metering is Hardware based so it can't
  • 00:12:35
    be ported to 40 series as is this is
  • 00:12:38
    where we get to Nvidia doing Nvidia
  • 00:12:40
    things and it's incorrect understanding
  • 00:12:43
    of present modon and how it works so
  • 00:12:44
    present Monon is a fully open source
  • 00:12:46
    utility it is maintained by Intel it can
  • 00:12:48
    be contributed to by basically anybody
  • 00:12:50
    other companies like AMD have worked
  • 00:12:52
    with Intel on maintaining and advancing
  • 00:12:55
    parts of present Monon so uh present
  • 00:12:57
    Monon measures performance it captures
  • 00:12:59
    data it's able to work at a lower level
  • 00:13:02
    with Windows and help to analyze the
  • 00:13:05
    frame time and the frame rate
  • 00:13:06
    performance animation error all kinds of
  • 00:13:08
    stuff it is uh the key backbone to most
  • 00:13:12
    of the performance analysis software
  • 00:13:13
    that's out there so even if you don't
  • 00:13:15
    use present Monon if you're an
  • 00:13:16
    Enthusiast who tries to measure your
  • 00:13:17
    performance you might use something else
  • 00:13:19
    that does now Intel's Tom Peterson has
  • 00:13:21
    joined us several times to explain
  • 00:13:23
    advancements to it and how it works it's
  • 00:13:25
    one of the most trusted testing
  • 00:13:27
    utilities for measuring component
  • 00:13:28
    performance and again its open-source
  • 00:13:30
    nature is critical to that especially
  • 00:13:33
    because it's maintained by Intel
  • 00:13:34
    unfortunately Nvidia has decided not to
  • 00:13:37
    contribute to and maintain present Manon
  • 00:13:39
    in a significant way and that puts us GN
  • 00:13:41
    in a weird position where we're acting
  • 00:13:43
    like a a marriage counselor we get to
  • 00:13:47
    talk to Intel on the presmont side and
  • 00:13:49
    Nvidia on the they don't understand how
  • 00:13:52
    it works side but they sure have some
  • 00:13:53
    opinions so Nvidia has instead taken an
  • 00:13:57
    ancient version of the present on core
  • 00:14:00
    inducted it into their frame view tool
  • 00:14:02
    which they build and maintain and it
  • 00:14:05
    then proceeded to make adjustments as
  • 00:14:07
    far as we can tell to that code but it
  • 00:14:10
    didn't do this in a way where it's
  • 00:14:11
    contributing to the public and current
  • 00:14:14
    branch of present Manon uh and is
  • 00:14:17
    instead maintaining its own thing as far
  • 00:14:19
    as we can tell it looks like Nvidia has
  • 00:14:22
    chosen not to contribute or work with
  • 00:14:24
    Intel on present Manon in a way that
  • 00:14:27
    would uh reflect the changes Nvidia is
  • 00:14:29
    trying to push in this section we're
  • 00:14:32
    getting into now this has gotten messy
  • 00:14:34
    because this results in bifurcation of
  • 00:14:36
    the present mon code into multiple
  • 00:14:39
    software solutions that might be on
  • 00:14:40
    different versions that are measuring
  • 00:14:42
    things differently this is a big problem
  • 00:14:44
    and Nvidia is sort of at the heart of it
  • 00:14:46
    so things like capf frame X or uh rtss
  • 00:14:50
    tools like this uh Hardware info they
  • 00:14:52
    use present Monon for the frame rate
  • 00:14:55
    analysis now we use just the present
  • 00:14:57
    Monon actual sort of core code from
  • 00:14:59
    present mon itself where we use uh
  • 00:15:01
    command line for everything we do but
  • 00:15:03
    recently as an example we introduced Jay
  • 00:15:05
    Jay's two cents to the present man
  • 00:15:07
    Graphics user interface and he's been
  • 00:15:10
    doing great with it so that's kind of
  • 00:15:11
    the core versus everything else that's
  • 00:15:12
    using it related to flip metering Nvidia
  • 00:15:16
    has pushed hard for a switch from
  • 00:15:17
    present manon's Ms between presents
  • 00:15:19
    metric to the MS between display change
  • 00:15:23
    metric while we understand why they're
  • 00:15:25
    trying to do this the premise we think
  • 00:15:27
    is flawed and that's because present man
  • 00:15:29
    already switched away from Ms between
  • 00:15:32
    presents but Nvidia pulled us aside at
  • 00:15:35
    CES and told us why it's time to switch
  • 00:15:37
    away from Ms between presents and then
  • 00:15:39
    it had a massive section of its review
  • 00:15:41
    guide about why it's time to switch away
  • 00:15:43
    from Ms between presents or milliseconds
  • 00:15:45
    between presents so again this kind of
  • 00:15:47
    goes back to that already happened but
  • 00:15:51
    Nvidia is using old code and so it's
  • 00:15:54
    sort of like being the last guy in town
  • 00:15:57
    using candle light walking outside and
  • 00:15:59
    saying I have a great idea what if we
  • 00:16:01
    switch to electricity we think nvidia's
  • 00:16:03
    frame view software is most likely based
  • 00:16:05
    on preson 1.10.0 from February 2024
  • 00:16:09
    based on inspecting the results output
  • 00:16:11
    the current version of present Monon
  • 00:16:12
    2.3.0 no longer supports either of the
  • 00:16:15
    metrics NV video is talking about under
  • 00:16:17
    those names it instead switched to a
  • 00:16:19
    frame time metric based on CPU busy and
  • 00:16:22
    CPU weight which Nvidia says is also
  • 00:16:25
    inadequate uh but Which Intel began
  • 00:16:28
    talking about around a year year ago Ms
  • 00:16:29
    between presents is an ancient metric
  • 00:16:31
    inherited from fraps which was basically
  • 00:16:34
    the only frame time tracking tool around
  • 00:16:36
    when present M began development it
  • 00:16:38
    tracks time between present calls which
  • 00:16:40
    tell the graphics engine to finish up
  • 00:16:42
    and put the frame on the screen but it
  • 00:16:44
    doesn't reflect animation time or time
  • 00:16:47
    on screen despite nvidia's decision to
  • 00:16:49
    not participate in working on the tool
  • 00:16:52
    that everybody else does and instead
  • 00:16:53
    build its own outdated toys and its own
  • 00:16:55
    playground uh it does have one good
  • 00:16:57
    point in all of this which is that the
  • 00:16:59
    valid part is uh situations such as MFG
  • 00:17:03
    back to what we're talking about today
  • 00:17:05
    have a problem here and the problem is
  • 00:17:07
    there's external frame metering very
  • 00:17:09
    late in the pipeline that may make the
  • 00:17:11
    onscreen frame rate smoother than Ms
  • 00:17:13
    between presents would imply but we
  • 00:17:15
    don't think Ms between display change is
  • 00:17:17
    sufficient either and that's what Nvidia
  • 00:17:18
    is saying everyone should use
  • 00:17:20
    milliseconds between display change also
  • 00:17:22
    doesn't track animation time Intel
  • 00:17:24
    defines it as quote how long the
  • 00:17:26
    previous frame was displayed before this
  • 00:17:28
    present was displayed in milliseconds
  • 00:17:31
    versus Ms between presents which was
  • 00:17:33
    quote the time between this present call
  • 00:17:36
    and the previous one in milliseconds end
  • 00:17:38
    quote in contrast the current present
  • 00:17:40
    mon frame time metric is the best
  • 00:17:42
    estimate of the time between the game
  • 00:17:44
    taking snapshots of its physics engine
  • 00:17:46
    however milliseconds between display
  • 00:17:47
    changes metric does take into account
  • 00:17:49
    flip metering for nvidia's MFG solution
  • 00:17:53
    and so it makes sense to use it
  • 00:17:54
    specifically for tracking nvidia's frame
  • 00:17:57
    generation and we mentioned animation
  • 00:17:59
    error which we've talked about again
  • 00:18:00
    with Tom Peterson in the past and this
  • 00:18:02
    is important just because frames are
  • 00:18:04
    delivered at an even Pace doesn't mean
  • 00:18:07
    they look good and it doesn't mean that
  • 00:18:09
    things being shown in those frames have
  • 00:18:11
    been animated at a consistent and even
  • 00:18:14
    Pace alongside the pacing of those
  • 00:18:16
    frames and animation times and flip
  • 00:18:18
    times have to be consistent this was a
  • 00:18:20
    problem with SLI and Crossfire where
  • 00:18:22
    frames could be delivered at a constant
  • 00:18:24
    rate but could appear stuttery anyway
  • 00:18:27
    because animation timing didn't line up
  • 00:18:29
    with the flip timing ideally in the long
  • 00:18:31
    term we'd move towards using FPS as a
  • 00:18:33
    metric for raw horsepower and animation
  • 00:18:36
    error as a metric for smoothness so to
  • 00:18:39
    speak but that's outside the scope of
  • 00:18:40
    this piece so there's a lot of technical
  • 00:18:42
    background on measurement tools and how
  • 00:18:44
    important they are for all these new
  • 00:18:46
    technologies coming out uh but to put a
  • 00:18:48
    very fine point on it for okay so what
  • 00:18:51
    what does all of this mean uh the so
  • 00:18:53
    what is NVIDIA is being a dick and
  • 00:18:56
    specifically doing so about present
  • 00:18:58
    Monon and open source tool and it's then
  • 00:19:00
    claiming that quote frame view has
  • 00:19:02
    solved a certain bugs to make
  • 00:19:04
    milliseconds between display change
  • 00:19:05
    accurate and more scenarios end quote
  • 00:19:07
    without communicating those changes to
  • 00:19:10
    the developers when we talked to Intel
  • 00:19:12
    about this relating to present Manon
  • 00:19:14
    they basically said
  • 00:19:16
    well Nvidia can talk to us if they want
  • 00:19:19
    to but somehow we're filtering it so
  • 00:19:21
    that puts us in a situation where in
  • 00:19:23
    order to test MFG and have any sort of
  • 00:19:26
    hopefully accuracy to it uh we have to
  • 00:19:29
    use nvidia's version of an open- Source
  • 00:19:32
    tool that Nvidia has closed off to
  • 00:19:34
    measure nvidia's Hardware with nvidia's
  • 00:19:37
    new technologies with supposedly bug
  • 00:19:40
    fixes that Nvidia has not explicitly
  • 00:19:43
    disclosed you can see why these are a
  • 00:19:45
    lot of problems for measuring an Nvidia
  • 00:19:48
    product it calls into question the
  • 00:19:49
    entire premise of the metrics uh because
  • 00:19:51
    there's no external way to really
  • 00:19:53
    thoroughly inspect what it is NVIDIA has
  • 00:19:55
    been doing behind the scenes now this
  • 00:19:57
    isn't illegal under zon's current
  • 00:19:59
    license but it is rude AMD is even
  • 00:20:02
    currently collaborating to support
  • 00:20:03
    tracking fluid motion frames but Nvidia
  • 00:20:05
    is not so anyway let's get back into MFG
  • 00:20:08
    and dlss now that we've gotten all the
  • 00:20:10
    measurement discussion this is all
  • 00:20:11
    really important because measuring these
  • 00:20:14
    fake frames is uh not necessarily
  • 00:20:17
    trivial and it requires some real
  • 00:20:19
    serious planning on the tools being used
  • 00:20:21
    so uh anyway we talk about image
  • 00:20:23
    enhancements next Nvidia is switching
  • 00:20:25
    away from convolutional neural networks
  • 00:20:26
    to a Transformer model for d s rolling
  • 00:20:29
    out as a beta this change affects all
  • 00:20:31
    RTX cards the gist is that the new model
  • 00:20:34
    should have a better understanding of
  • 00:20:35
    which areas of the screen are important
  • 00:20:37
    and therefore has a quote greater
  • 00:20:40
    stability reduced ghosting higher detail
  • 00:20:42
    in motion and smoother edges in a scene
  • 00:20:44
    and quote this is an image quality
  • 00:20:46
    improvement not a performance
  • 00:20:47
    Improvement and may actually worsen
  • 00:20:49
    performance a bit Nvidia notes that
  • 00:20:51
    quote while Ada and Blackwell both
  • 00:20:53
    benefit from this new model the
  • 00:20:55
    Transformer model is often more
  • 00:20:56
    performant on Blackwell end quote and
  • 00:20:58
    we'll show our own results in a few
  • 00:21:00
    minutes here for each of the DLS has
  • 00:21:02
    four features we've discussed the new
  • 00:21:03
    frame generation model MFG the
  • 00:21:05
    Transformer model for DLS SSR and RR
  • 00:21:08
    there are now driver level overrides for
  • 00:21:11
    example with a game that supported the
  • 00:21:12
    original FG and a card that supports MFG
  • 00:21:15
    you could use the driver to force MFG
  • 00:21:18
    this is a whit list system so not every
  • 00:21:21
    game will support overriding and the
  • 00:21:23
    ones that do won't necessarily have
  • 00:21:25
    every dss4 feature available we don't
  • 00:21:27
    expect this to get a ton of VI the
  • 00:21:29
    interface is currently confusing and
  • 00:21:30
    we'd assume that most developers is
  • 00:21:32
    willing to go through the hassle of
  • 00:21:33
    getting whitelisted for overrides will
  • 00:21:35
    just update their games there are plenty
  • 00:21:37
    of other new features that were
  • 00:21:38
    announced alongside the 50 Series but
  • 00:21:39
    we're just talking about dlss here
  • 00:21:41
    reflex is technically not dlss but it is
  • 00:21:44
    a dependency so we're throwing it in
  • 00:21:46
    reflex's new feature is frame warp which
  • 00:21:49
    is similar to the reprojection tech
  • 00:21:51
    that's been used in VR it checks Mouse
  • 00:21:53
    input at the last possible moment and if
  • 00:21:55
    it detects an update it shifts the
  • 00:21:56
    camera over to approximate the correct
  • 00:21:59
    movement the unrendered pixels are then
  • 00:22:01
    painted in whether or not this
  • 00:22:03
    represents a literal reduction in
  • 00:22:04
    latency is arguable but it should feel
  • 00:22:06
    like it time to get into performance
  • 00:22:08
    testing this is really complicated for
  • 00:22:09
    all of the wrong reasons so uh we have a
  • 00:22:12
    couple groups of comparisons here our
  • 00:22:13
    first and primary objective is going to
  • 00:22:15
    be to compare the old and the new models
  • 00:22:17
    for frame generation to each other and
  • 00:22:19
    so what we're looking at is what's the
  • 00:22:20
    difference between the two models we're
  • 00:22:22
    not looking at the 490 versus the 590 we
  • 00:22:24
    have both sets of results but it's not a
  • 00:22:26
    comparison of cards it's a comparison of
  • 00:22:28
    model mod running on cards for frame
  • 00:22:30
    generation there's a couple challenges
  • 00:22:32
    of analyzing so-called fake frames and
  • 00:22:35
    one of them is that there isn't a clear
  • 00:22:37
    amount of generated frames necessarily
  • 00:22:40
    so in other words it's not necessarily
  • 00:22:42
    consistent how many generated frames
  • 00:22:43
    there are and that poses a challenge
  • 00:22:46
    another challenge is that although we
  • 00:22:47
    will have a control on the charts for I
  • 00:22:50
    don't raw frames pure frames the purest
  • 00:22:53
    of frames with no artificial pixels in
  • 00:22:55
    them whatever you call those normal fr
  • 00:22:58
    frames we're going to have a control for
  • 00:23:00
    those but generated frames by the very
  • 00:23:03
    nature of them are not alike for like
  • 00:23:05
    comparison because they're not the same
  • 00:23:07
    frames they contain generated data and
  • 00:23:12
    so the whole thing is just it's it makes
  • 00:23:14
    it challenging to do true head-to-head
  • 00:23:16
    comparisons which is why we have such a
  • 00:23:18
    problem with Nvidia that marketing
  • 00:23:19
    charge that's a different discussion all
  • 00:23:20
    right so our test bench is detailed on
  • 00:23:22
    the website it's free of thirdparty ads
  • 00:23:24
    and it includes the hardware
  • 00:23:25
    configuration you can check that out
  • 00:23:27
    scroll through it see the hardware if
  • 00:23:28
    you want to see it and another important
  • 00:23:30
    thing is that we're not covering image
  • 00:23:32
    quality in this piece that'll come later
  • 00:23:34
    but we'll go over performance numbers
  • 00:23:36
    today image quality is an entirely
  • 00:23:38
    different Deep dive Nvidia provided six
  • 00:23:40
    games to the Press with the new DLS S4
  • 00:23:42
    features available we dug through them
  • 00:23:44
    to find some way to isolate the features
  • 00:23:46
    for testing four of the games had native
  • 00:23:48
    feature implementations and two were
  • 00:23:50
    whitelisted for driver overrides as
  • 00:23:52
    mentioned earlier we were forced to use
  • 00:23:54
    frame view for all benchmarks involving
  • 00:23:56
    frame generation for the purposes of
  • 00:23:58
    these tests we're including generated
  • 00:24:00
    frames in our averages our first task is
  • 00:24:02
    to compare the old frame gen model
  • 00:24:04
    performance versus the new frame gen
  • 00:24:05
    model performance on the 490 and 590 the
  • 00:24:08
    company hasn't claimed that image
  • 00:24:09
    quality will change or improve with the
  • 00:24:11
    new method the only two games available
  • 00:24:13
    to us with both models for Marvel Rivals
  • 00:24:15
    and Dragon Age Veil guard both through
  • 00:24:17
    driver overrides but we were also able
  • 00:24:20
    to make a usable comparison with
  • 00:24:21
    Hogwarts Legacy we're using the games
  • 00:24:23
    that work with it for the charts coming
  • 00:24:24
    up focus on the red versus the orange
  • 00:24:26
    bars that's where we look at the two mod
  • 00:24:28
    the blue bar which is the uh control
  • 00:24:31
    Baseline no frame generation it's kind
  • 00:24:32
    of questionable whether that can be
  • 00:24:34
    directly compared to the FG models
  • 00:24:36
    because of the differences we've already
  • 00:24:37
    talked about there's one other confusing
  • 00:24:38
    aspect as well if we want to compare the
  • 00:24:40
    impact of overhead on real frame
  • 00:24:43
    rendering as well then we need to know
  • 00:24:45
    how many fake frames were generated in
  • 00:24:46
    the fake plus real frame rate to then
  • 00:24:49
    compare against the real frames in the
  • 00:24:51
    real only frame non-generated rendering
  • 00:24:53
    in other words if the frames are half
  • 00:24:55
    fake then we need to real cut the fake
  • 00:24:58
    frames in real half because the half
  • 00:25:01
    fake real having would real have the
  • 00:25:03
    percent uplift caused by the half fake
  • 00:25:06
    frames but it's not clear if you can
  • 00:25:09
    just divide by two because MFG generates
  • 00:25:11
    up to three fake frames that are fully
  • 00:25:14
    artificial for every one real frame but
  • 00:25:17
    then it's up two and so it's not as easy
  • 00:25:19
    as maybe just dividing and anyway this
  • 00:25:21
    is getting way too complicated you can
  • 00:25:23
    see exactly what the issue is with all
  • 00:25:24
    of this we're going to put numbers on
  • 00:25:26
    the screen and then our solution to the
  • 00:25:27
    complication is we're going to read the
  • 00:25:30
    numbers that are on the screen and do
  • 00:25:32
    percent comparison between the two
  • 00:25:34
    models that is the simplest thing we can
  • 00:25:36
    do you get the idea it's a very tricky
  • 00:25:37
    subject this benefits Nvidia with it
  • 00:25:40
    being tricky and uh the whole thing is
  • 00:25:43
    kind of it's challenging to say how
  • 00:25:46
    precisely like for like anything is but
  • 00:25:48
    that is the nature of artificially
  • 00:25:50
    generated data so let's get into some
  • 00:25:52
    charts that are real of at least half
  • 00:25:56
    fake data but that's real so here they
  • 00:26:00
    are here's Dragon Age Veil guard at 4K
  • 00:26:02
    with Max settings we manually increased
  • 00:26:03
    it above the highest preset for this
  • 00:26:05
    dlss super resolution or Sr and AA or
  • 00:26:08
    anti-aliasing were not used for these
  • 00:26:10
    tests and reflex was enabled even for
  • 00:26:12
    the control runs without frame
  • 00:26:13
    generation each card should be compared
  • 00:26:15
    against itself here so we'll start with
  • 00:26:16
    the 490 the new model with f2x on the
  • 00:26:19
    490 ran at 107 FPS average against the
  • 00:26:22
    old models result of 100.7 FPS average
  • 00:26:25
    an increase in fake plus real frame
  • 00:26:27
    perceived fake Plus real performance of
  • 00:26:29
    6.26 probably real per. if we were to
  • 00:26:31
    filter out the fake frames to only real
  • 00:26:33
    frames to real evaluate the fake
  • 00:26:35
    overhead then that percentage comes down
  • 00:26:37
    hopefully you're following the RTX 590
  • 00:26:39
    ran at 143.5 FPS average on the new
  • 00:26:42
    model and 136.6 on the old model an
  • 00:26:45
    improvement in the as captured result of
  • 00:26:47
    5.1% uplift in perceived performance
  • 00:26:50
    again there's more Nuance if you filter
  • 00:26:52
    out the artificial frames Hogwarts
  • 00:26:53
    legacy's up next this one didn't have a
  • 00:26:55
    driver level override for the frame
  • 00:26:57
    generation model so we had to control
  • 00:26:59
    that variable by switching back and
  • 00:27:01
    forth for the Press build otherwise the
  • 00:27:03
    results would be invalid but
  • 00:27:04
    complicating it more Ray tracing
  • 00:27:06
    features were updated in the Press build
  • 00:27:09
    and that would make it invalid so we
  • 00:27:11
    disabled all RT effects for these tests
  • 00:27:14
    we first compared the public and the
  • 00:27:15
    Press versions of the game with frame
  • 00:27:17
    generation off to ensure that the
  • 00:27:19
    comparison even made any sense and was a
  • 00:27:22
    valid comparison disabling RT
  • 00:27:23
    effectively equalized the performance
  • 00:27:25
    between the game versions 107.3 to 10
  • 00:27:28
    6.9 is with an error as you see in this
  • 00:27:29
    chart and
  • 00:27:32
    151.90 is also with an error that makes
  • 00:27:34
    these valid between the versions under
  • 00:27:36
    these specific conditions that brings us
  • 00:27:38
    to this chart in Hogwarts Legacy the 490
  • 00:27:40
    had a 14% uplift in perceived
  • 00:27:42
    performance from 153 fps to 175 FPS
  • 00:27:46
    average with the new model while the
  • 00:27:48
    5090 had a 10% uplift from 224 to 247
  • 00:27:53
    FPS average working backwards based on
  • 00:27:55
    the output frame rates we can deduce
  • 00:27:56
    that the new frame generation model
  • 00:27:58
    model is easier on the GPU this allows
  • 00:28:00
    more resources to be dedicated to
  • 00:28:01
    rendering it enables a higher input
  • 00:28:03
    frame rate and that then comes back
  • 00:28:05
    around to benefiting frame generation
  • 00:28:08
    the other piece of the puzzle will be
  • 00:28:09
    determining whether subjective image
  • 00:28:11
    quality has changed in a later video but
  • 00:28:14
    from a performance perspective today
  • 00:28:15
    switching away from OFA is better the
  • 00:28:18
    second feature we tested in isolation
  • 00:28:19
    was the old convolutional neural network
  • 00:28:21
    model versus the new Transformer model
  • 00:28:24
    this affects both dlss super resolution
  • 00:28:26
    and dlss Ray reconstruction we were able
  • 00:28:28
    to test Sr in isolation but RR Ray
  • 00:28:32
    reconstruction requires Sr to be enabled
  • 00:28:35
    again an infinite list of nested
  • 00:28:38
    variables first up is Veil guard again
  • 00:28:40
    and using the driver override feature
  • 00:28:42
    for AB testing Ray reconstruction isn't
  • 00:28:45
    available as a feature in this game so
  • 00:28:46
    this is purely a super resolution test
  • 00:28:48
    we used the dlss SR quality preset which
  • 00:28:51
    given the 4K output resolution means a
  • 00:28:54
    2560 X 1440 render resolution on the
  • 00:28:56
    4090 performance softw using the new
  • 00:28:58
    Transformer model versus the older CNN
  • 00:29:00
    model with the older model retaining a
  • 00:29:03
    5% performance Advantage the No dlss
  • 00:29:06
    Control result is included for context
  • 00:29:08
    but that's not the comparison we're
  • 00:29:10
    focusing on here on the 90 the older CNN
  • 00:29:12
    model had a similar performance
  • 00:29:14
    advantage of 4% averaging 105 FPS versus
  • 00:29:17
    101 with the Transformer model the
  • 00:29:19
    difference between the performance
  • 00:29:20
    degradation on the two cards is
  • 00:29:21
    minuscule here but these results would
  • 00:29:23
    technically align with nvidia's Claim
  • 00:29:24
    about the Transformer model being quote
  • 00:29:26
    more performant on Blackwell end quote
  • 00:29:29
    the differences are too small here to
  • 00:29:31
    make that a firm conclusion what we can
  • 00:29:33
    say with certainty is that the new model
  • 00:29:35
    is a little lower performance than the
  • 00:29:37
    old one with Veil guard cyberpunk 2077
  • 00:29:40
    crashing Liberty is up next when it
  • 00:29:42
    works it has Ray reconstruction as a
  • 00:29:44
    feature so we were able to test both
  • 00:29:46
    dlss Sr alone and dlss Sr and RR in
  • 00:29:50
    combination when it worked as noted
  • 00:29:52
    during our dlss 3.5 coverage it's
  • 00:29:54
    possible for R Rec construction to raise
  • 00:29:56
    or lower performance depending on the
  • 00:29:58
    systems that it replaces in a given game
  • 00:30:01
    however it continues to have no
  • 00:30:03
    significant performance impact in
  • 00:30:04
    cyberpunk we recorded results without
  • 00:30:06
    Ray reconstruction but we cut them from
  • 00:30:08
    these charts because it made no
  • 00:30:10
    difference the losses aren't huge the
  • 00:30:12
    490 with dlss and Ray reconstruction
  • 00:30:14
    dropped from 71 FPS with CNN to 68 with
  • 00:30:18
    the Transformer Model A 5% Advantage for
  • 00:30:21
    the old model in terms of performance on
  • 00:30:23
    the 5090 switching to the new model
  • 00:30:24
    dropped average FPS from 92 to 86 7 %
  • 00:30:28
    performance Advantage for the older
  • 00:30:29
    model on this card anecdotally we
  • 00:30:31
    noticed that the newer model reduces
  • 00:30:33
    ghosting and visibly improves the
  • 00:30:36
    appearance of animated textures and
  • 00:30:37
    cyberpunk so this level of performance
  • 00:30:39
    loss is likely an acceptable trade-off
  • 00:30:42
    that comes back to a future image
  • 00:30:43
    quality piece though because the new DLS
  • 00:30:45
    ssrr model comes with a performance hit
  • 00:30:48
    unlike the straight upgrade to FG we
  • 00:30:50
    have enough acronyms for you we expect
  • 00:30:52
    many games updated to support the model
  • 00:30:54
    will take the same approach as cyberp
  • 00:30:56
    Punk hopefully without the crashing and
  • 00:30:57
    offer a choice between CNN and
  • 00:30:59
    Transformer in settings hopefully those
  • 00:31:01
    toggles actually work also unlike
  • 00:31:03
    cyberpunk because they remain
  • 00:31:04
    problematic in this game we always
  • 00:31:06
    relaunch games between settings changes
  • 00:31:08
    of any kind but we noticed that
  • 00:31:09
    relaunching cyberpunk with our specific
  • 00:31:11
    settings would switch to the Transformer
  • 00:31:13
    model invisibly while claiming to still
  • 00:31:16
    use CNN in the menu we were forced to
  • 00:31:18
    toggle CNN off and on from inside the
  • 00:31:22
    game in order to test it reliably all of
  • 00:31:25
    this was observed in the public 2.21
  • 00:31:27
    steam release this is an extremely
  • 00:31:29
    subtle bug it wastes a huge amount of
  • 00:31:31
    time and it could invalidate data so to
  • 00:31:33
    be as efficient as possible we'll just
  • 00:31:35
    recycle our cyberpunk rant from 2 years
  • 00:31:37
    ago because it's still valid today
  • 00:31:39
    here's a clip also if cdpr is listening
  • 00:31:42
    please fix the settings menu that we
  • 00:31:45
    talked about earlier individual settings
  • 00:31:47
    get reset constantly for inscrutable
  • 00:31:49
    reasons we've had situations where we've
  • 00:31:50
    applied settings click run Benchmark in
  • 00:31:52
    the settings menu then the settings at
  • 00:31:54
    the end of the bench were different from
  • 00:31:55
    the ones that we'd set at the start it's
  • 00:31:57
    extremely diff difficult to have
  • 00:31:58
    confidence that the settings are ever
  • 00:32:00
    what they say they are our final
  • 00:32:01
    performance comparisons are frame pacing
  • 00:32:03
    with frame generation enabled this comes
  • 00:32:05
    with two major caveats first we're using
  • 00:32:07
    frame view for logging so the plotted
  • 00:32:09
    frame times are milliseconds between
  • 00:32:10
    display change second we don't have a
  • 00:32:12
    way to toggle frame metering on the 590
  • 00:32:15
    so we're forced to make a general
  • 00:32:17
    comparison between fg2 X behavior on the
  • 00:32:19
    490 without frame metering versus the
  • 00:32:21
    590 with frame metering ideally we'd run
  • 00:32:23
    the FG 4X on the 90 alone with versus
  • 00:32:27
    without frame metering the new frame gen
  • 00:32:29
    model was used for these comparisons so
  • 00:32:32
    here's Dragon Age again as always making
  • 00:32:34
    a comparative plot is Complicated by the
  • 00:32:36
    fact that the higher performance card
  • 00:32:37
    renders more frames the 490s plot shows
  • 00:32:39
    the kind of up and down variation
  • 00:32:41
    between every frame that might be
  • 00:32:43
    expected from frame generation with the
  • 00:32:45
    majority of the pass oscillating up and
  • 00:32:46
    down by 1 to 2 milliseconds on each
  • 00:32:48
    individual frame that may not be enough
  • 00:32:50
    to cause a visible problem in this title
  • 00:32:53
    but it's a good illustration of the
  • 00:32:54
    problem that Nvidia is trying to address
  • 00:32:56
    the 590 actually has had visibly wider
  • 00:32:58
    deviations from the average on the whole
  • 00:33:00
    but the up and down deviations didn't
  • 00:33:02
    occur between every single consecutive
  • 00:33:04
    frame we'd want to see much flatter
  • 00:33:06
    lines even still ideally most people
  • 00:33:08
    notice excursions at 8 to 12
  • 00:33:10
    milliseconds were not there with this
  • 00:33:11
    chart but based on prior interviews
  • 00:33:13
    we've done that seems be where people
  • 00:33:14
    notice it you'd still want a flatter
  • 00:33:16
    line ideally though Hogwarts Legacy
  • 00:33:18
    showed results closer to what Nvidia
  • 00:33:19
    promised with frame metering the 490s
  • 00:33:21
    still exhibit some up and down Behavior
  • 00:33:23
    between individual frames as real and
  • 00:33:25
    fake frames alternate while the 509s
  • 00:33:27
    frame are clearly more stable the game
  • 00:33:29
    also renders at a very high frame rate
  • 00:33:31
    though so the actual time Delta in
  • 00:33:33
    milliseconds between frames is minuscule
  • 00:33:35
    even on the 490 the 90 also doesn't
  • 00:33:38
    completely eliminate the occasional
  • 00:33:39
    frame time Spike we still see those in
  • 00:33:41
    this chart with a particularly large one
  • 00:33:43
    shown at the end of this past so we
  • 00:33:44
    arrive here this is one of the videos
  • 00:33:46
    where jumping to the conclusion is going
  • 00:33:47
    to actually cut out a lot of really
  • 00:33:49
    important stuff like measurement
  • 00:33:51
    methodology and all of these different
  • 00:33:52
    things that nvidia's kind of they've got
  • 00:33:54
    their way they want you to do it and we
  • 00:33:57
    just think that it's kind of a
  • 00:33:59
    reason because look they could
  • 00:34:00
    contribute to present Manon like
  • 00:34:02
    everyone else and we could make it a lot
  • 00:34:04
    less complicated but anyway breaking
  • 00:34:06
    down dss4 into its individual components
  • 00:34:08
    the new frame generation model does
  • 00:34:10
    perform better broadly speaking and
  • 00:34:12
    there shouldn't be a trade-off inv
  • 00:34:13
    visual Fidelity but we haven't fully
  • 00:34:15
    confirmed that yet we've just kind of
  • 00:34:16
    anecdotally looked at it we have a
  • 00:34:18
    separate image quality comparison that
  • 00:34:19
    we're going to be working on we haven't
  • 00:34:20
    started yet I think we're starting
  • 00:34:21
    tomorrow probably and um that's where
  • 00:34:23
    we're going to look at the image quality
  • 00:34:24
    tradeoff so this is a very big topic and
  • 00:34:27
    has a lot of composite Parts we're going
  • 00:34:29
    to tackle it in pieces and then uh we'll
  • 00:34:31
    be able to form a full picture from all
  • 00:34:33
    that latency is another one of those
  • 00:34:34
    things that we're going to be looking at
  • 00:34:35
    as well in relation to all of this so
  • 00:34:38
    this is just to be clear this is only a
  • 00:34:39
    part of the story which is intentional
  • 00:34:42
    just because we want to really go in
  • 00:34:43
    depth with every part of it as much as
  • 00:34:45
    we can but anyway the updated model also
  • 00:34:47
    theoretically makes single frame
  • 00:34:48
    generation possible uh on tensor
  • 00:34:51
    equipped RTX card and Nvidia has at this
  • 00:34:55
    point made no indication that it plans
  • 00:34:57
    to support that but in theory it could
  • 00:34:59
    be supported on things that are not
  • 00:35:00
    currently officially supported the new
  • 00:35:02
    Transformer model for dlss Sr and RR
  • 00:35:05
    performs worse than the older CNN model
  • 00:35:07
    but it should also look better and our
  • 00:35:09
    experience so far anecdotally suggests
  • 00:35:12
    that it is worth it for that we'll
  • 00:35:13
    follow up with that frame pacing with
  • 00:35:15
    frame generation enabled should be
  • 00:35:17
    improved versus the 40 series but this
  • 00:35:19
    would mostly apply to MFG and the 40
  • 00:35:22
    series can't do MFG still our extremely
  • 00:35:24
    limited testing on that aspect does show
  • 00:35:27
    that frame metering has an effect we'd
  • 00:35:29
    also like to really drive home this one
  • 00:35:30
    key point because it keeps coming up not
  • 00:35:32
    just online but uh even in discussions
  • 00:35:35
    in the office where other people who
  • 00:35:36
    aren't part of this project are kind of
  • 00:35:37
    asking about it the same concept keeps
  • 00:35:40
    coming up which is that uh frame gen is
  • 00:35:43
    really to make already acceptable frame
  • 00:35:46
    rates into so-called smooth frame rates
  • 00:35:49
    it's it's more of a smoothing thing than
  • 00:35:51
    it is a corrective thing so kind of
  • 00:35:53
    break that habit that uh Jensen Juan has
  • 00:35:57
    reinforced on stage that it fixes a
  • 00:36:01
    problem because what it's really doing
  • 00:36:04
    is helping something that's already
  • 00:36:05
    acceptable uh just be a little bit
  • 00:36:08
    smoother but it's not going to fix a
  • 00:36:11
    25fps sort of native rendering rate and
  • 00:36:16
    turn that into something that feels like
  • 00:36:19
    it's truly 50 or something like that
  • 00:36:20
    you're not going to get that experience
  • 00:36:22
    in the same way as a native 50 would in
  • 00:36:24
    that example so the place to use this
  • 00:36:25
    would be things like 4K 240 ol ads that
  • 00:36:28
    have been coming out recently sometimes
  • 00:36:30
    Nvidia seems aware of this and other
  • 00:36:31
    times it decides that it's not
  • 00:36:33
    advantageous to be aware of it like when
  • 00:36:35
    it says that the 5070 is equivalent to a
  • 00:36:38
    490 because of MFG even though in that
  • 00:36:41
    presentation in one sentence he makes it
  • 00:36:42
    very clear where he's like 570 $550
  • 00:36:45
    equals a 490 all because of AI 490
  • 00:36:48
    performance at $549 impossible without
  • 00:36:51
    artificial intelligence and then
  • 00:36:51
    literally the next sentence he drops the
  • 00:36:53
    because of AI part and it just kind of
  • 00:36:55
    morphs into 570 equals 4 90 see previous
  • 00:36:59
    sentence for the really important caveat
  • 00:37:01
    if you have bad performance frame
  • 00:37:03
    generation May disappoint you dlss super
  • 00:37:06
    resolution is a more helpful option in
  • 00:37:08
    that situation where you can just modify
  • 00:37:10
    the actual render resolution and gain
  • 00:37:12
    some more frame rate through that method
  • 00:37:14
    so that's the way to think about this
  • 00:37:15
    stuff we think for now and we're really
  • 00:37:17
    just getting started in this it's fun
  • 00:37:19
    there's a lot of cool stuff that we get
  • 00:37:20
    to look at with this and uh part of that
  • 00:37:23
    is because Nvidia is inconsistent with
  • 00:37:26
    when it claims certain things
  • 00:37:28
    and we think that there may be strategic
  • 00:37:30
    reasons for that and uh it would benefit
  • 00:37:33
    them if it's a little confusing
  • 00:37:34
    sometimes it could also just be that
  • 00:37:37
    people are actually legitimately
  • 00:37:38
    confused about all this stuff that's
  • 00:37:39
    what we're trying to help with um we
  • 00:37:41
    have been ramping on this because it's a
  • 00:37:43
    really detailed topic I mean the review
  • 00:37:45
    guide is like over a 100 pages long and
  • 00:37:47
    we read that and we had a lot of
  • 00:37:49
    complaints about some of the stuff they
  • 00:37:50
    wrote in there we talked about a lot of
  • 00:37:51
    those today uh but even still it's
  • 00:37:53
    complicated so next things are image
  • 00:37:55
    quality we're going to work on that and
  • 00:37:57
    then uh I'd like to do latency I don't
  • 00:38:00
    know what our timeline is on latency but
  • 00:38:02
    IQ we're going to work on next so keep
  • 00:38:04
    an eye out for image quality comparisons
  • 00:38:06
    and otherwise lot of cool stuff coming
  • 00:38:07
    up we have some more GPU coverage to
  • 00:38:09
    kind of close us out with the ITX case
  • 00:38:11
    testing with the flowthrough design
  • 00:38:12
    that's going to be really fun and just
  • 00:38:13
    kind of a scientific experiment then we
  • 00:38:15
    also have some impact to CPU tower
  • 00:38:17
    cooling performance uh in a traditional
  • 00:38:19
    ATX case with that dual flow through
  • 00:38:20
    that's something I'm just personally
  • 00:38:22
    really excited about because it's a fun
  • 00:38:24
    test it's the kind of stuff that I love
  • 00:38:25
    working on just to see what it does uh
  • 00:38:27
    but beyond those we're probably going to
  • 00:38:29
    start looking back to pre-builts and uh
  • 00:38:32
    cases and things like that to get you
  • 00:38:33
    some more content diversity but uh
  • 00:38:35
    there's a lot of stuff for GPS right now
  • 00:38:37
    so stick with us through that and at the
  • 00:38:39
    very least we have some fun feature
  • 00:38:40
    tests check back for more subscribe as
  • 00:38:42
    always you can go to store. Gamers
  • 00:38:43
    access.net to help us out directly with
  • 00:38:45
    this testing and the uh amount of
  • 00:38:48
    additional information we try to provide
  • 00:38:49
    these days in the videos and you can go
  • 00:38:51
    to patreon.com Gamers if you want to
  • 00:38:53
    throw a few bucks our way to help fund
  • 00:38:55
    our operation thanks for watching we
  • 00:38:57
    we'll see you all next time
Tags
  • Nvidia
  • DLSS 4.0
  • MFG
  • Open source
  • Frame generation
  • Image quality
  • Performance
  • Transformer model
  • GPU
  • Testing